The state-of-the-art machine learning models can extract hundreds of high-level topics from large-scale social media corpus. Nevertheless, it remains a challenge for the users to interpret the comprehensive distribution of multiple topics in a space-time setting. In this paper, we present TopicFields, an interactive system to explore, aggregate, and visualize geo-tagged social media using hybrid topic models, scalar fields, and stream graphs. In the data processing stage, we apply two machine learning models Word2Vec and Inception-v3 to the data and address the relationships among the extracted topics by rearranging them via spectral ordering. In the visualization stage, we allow users to interactively select the preferred topics and alter the transfer function for visualizing the social media on a map with levels of detail. Our system, TopicFields, can efficiently estimate the kernel density distribution and visualize the scalar fields of the user-selected topics on a map on the GPU. In addition, we use temporal filters and stream graphs to enhance comprehensibility of the data over time. Here we present the system and its architecture that ingests geo-tagged Instagram and Twitter messages, extracts topics, hierarchically clusters, and facilitates their interactive visualization on a map. We demonstrate the effectiveness of TopicFields with several potential use cases. We envision our system will be useful for visual analytics of geo-tagged social media, tourism itinerary planning, and business intelligence.